GPT-6-Astra solved 49 of the 50 unpublished theoretical physics problems in the TheoPhysEval benchmark, scoring 98% on the track that allowed web searches. Its performance fell to 46 problems out of 50 without internet access, while Opus 5.5 scored 46 in both conditions. The benchmark, posted on arXiv on October 9, 2026, covers 11 subfields and was developed by 35 authors from institutions including ICTS and Google DeepMind. One problem remained unsolved by every model tested, and some human-written reference solutions were corrected. The run with web access cost $116.28, but the paper has not yet undergone peer review and does not measure the models’ ability to select or formulate new research questions.
Source: Read the original article
Disclaimer: this content is for information purposes only and is not financial advice. Cryptocurrencies are highly volatile: you may lose all of your capital. Always do your own research. Legal notice
